absl::Span<AbstractTensorHandle* const> inputs,
                               int input_index, bool use_function,
                               AbstractTensorHandle** numerical_grad) {
  vector<AbstractTensorHandle*> theta_inputs(inputs.size());
  for (int i{}; i < inputs.size(); ++i) {
    theta_inputs[i] = inputs[i];
  }

  AbstractTensorHandle* theta =

  ASSERT_EQ(read_message, message);
}

absl::Status RecordOperationWithNullGradientFunctionModel(
    AbstractContext* ctx, absl::Span<AbstractTensorHandle* const> inputs,
    absl::Span<AbstractTensorHandle*> outputs) {
  Tape tape(/*persistent=*/false);
  tape.Watch(inputs[0]);
  AbstractTensorHandle* neg_output;
  TF_RETURN_IF_ERROR(ops::Neg(ctx, inputs[0], &neg_output, "Neg"));
  tape.RecordOperation(inputs, {neg_output}, nullptr, "Neg");

// GraphContext and vice-versa).
class TracingTensorHandle : public AbstractTensorHandle {
 protected:
  explicit TracingTensorHandle(AbstractTensorHandleKind kind)
      : AbstractTensorHandle(kind) {}

 public:
  // For LLVM style RTTI.
  static bool classof(const AbstractTensorHandle* ptr) {
    return ptr->getKind() == kGraph || ptr->getKind() == kMlir;
  }
};

// Abstract interface to a Tensor handle in either tracing or immediate
// execution mode.
class AbstractTensorHandle : public core::RefCounted {
 protected:
  enum AbstractTensorHandleKind { kGraph, kMlir, kEager, kTfrt, kCustomDevice };
  explicit AbstractTensorHandle(AbstractTensorHandleKind kind) : kind_(kind) {}
  ~AbstractTensorHandle() override {}

 public:
  // Returns tensor dtype.

limitations under the License.
==============================================================================*/

#include "tensorflow/c/eager/abstract_tensor_handle.h"

namespace tensorflow {

std::string AbstractTensorHandle::DebugString() const {
  PartialTensorShape shape;
  absl::Status s = Shape(&shape);
  std::string shape_string;
  if (!s.ok()) {
    shape_string = "<error computing shape>";
  } else {

  // Converts a Gradient to a TapeTensor.
  TapeTensor TapeTensorFromGradient(AbstractTensorHandle* g) const override;

  void MarkAsResult(AbstractTensorHandle* gradient) const override;

  void DeleteGradient(AbstractTensorHandle* gradient) const override;

 private:
  // The context where the aggregation op `Add` is to be created.
  AbstractContext* ctx_;
};

// Records the inputs in the `ForwardOperation`.
absl::Status AddInput(AbstractOperation*, AbstractTensorHandle*,
                      ForwardOperation*);
absl::Status AddInputList(AbstractOperation*,
                          absl::Span<AbstractTensorHandle* const> inputs,
                          ForwardOperation*);

// Sets the attrs in the `ForwardOperation`.

 *
 * `numerical_grad` is the pointer to the AbstractTensorHandle* which will
 * hold the numerical gradient data at the end of the function.
 */
absl::Status CalcNumericalGrad(AbstractContext* ctx, Model forward,
                               absl::Span<AbstractTensorHandle* const> inputs,
                               int input_index, bool use_function,
                               AbstractTensorHandle** numerical_grad);

}  // namespace gradients

  // For LLVM style RTTI.
  static bool classof(const AbstractTensorHandle* ptr) {
    return ptr->getKind() == kEager || ptr->getKind() == kTfrt;
  }

 protected:
  explicit ImmediateExecutionTensorHandle(AbstractTensorHandleKind kind)
      : AbstractTensorHandle(kind) {}
  ~ImmediateExecutionTensorHandle() override {}
};

namespace internal {

  }

  absl::Status AddInput(AbstractTensorHandle* input) override {
    GraphTensor* t = dyn_cast<GraphTensor>(input);
    if (!t) {
      return tensorflow::errors::InvalidArgument(
          "Unable to cast input to GraphTensor");
    }
    TF_AddInput(op_.get(), t->output_);
    return absl::OkStatus();
  }
  absl::Status AddInputList(
      absl::Span<AbstractTensorHandle* const> inputs) override {